Enhanced Exploration of Neural Network Models for Indoor Human Monitoring

Indoor human monitoring can enable or enhance a wide range of applications, from medical to security and home or building automation. For effective ubiquitous deployment, the monitoring system should be easy to install and unobtrusive, reliable, low cost, tagless, and privacy-aware. Long-range capacitive sensors are good candidates, but they can be susceptible to environmental electromagnetic noise and require special signal processing. Neural networks (NNs), especially 1D convolutional neural networks (1D-CNNs), excel at extracting information and rejecting noise, but they lose important relationships in max/average pooling operations. We investigate the performance of NN architectures for time series analysis without this shortcoming, the capsule networks that use dynamic routing, and the temporal convolutional networks (TCNs) that use dilated convolutions to preserve input resolution across layers and extend their receptive field with fewer layers. The networks are optimized for both inference accuracy and resource consumption using two independent state-of-the-art methods, neural architecture search and knowledge distillation. Experimental results show that the TCN architecture performs the best, achieving 12.7% lower inference loss with 73.3% less resource consumption than the best 1D-CNN when processing noisy capacitive sensor data for indoor human localization and tracking.